「構造不変の定理」とそれに基づく音声の構造的表象： 心理学・言語学・情報学・物理学・複雑系の接点
Structural representation of speech based on a theorem of the invariant structure: Speech representation by integrating psychology, linguistics, informatics, physics, and complexity

峯松信明 ( 東京大学大学院新領域創成科学研究科)
Professor Nobuaki Minematsu, Graduate School of Frontier Sciences, The University of Tokyo

概要：
『音声知覚の容易性と音声物理の多様性』
音声研究者を長年悩ませ続けた古典的問題である。音声科学は，音素に対する音響不変量の不在から，調音運動にその解を求め，音声工学は，膨大なるデータ収集と統計的モデリングにその解を求めている。問題は解かれたのか？否である。本講演ではこの問題を，「構造不変の定理」と呼ぶ数学定理の上で，古典的な言語学の議論（構造音韻論）を物理実装する形で解く。その結果，音声ゲシュタルトとでも呼ぶべき音声の構造的表象が導かれる。その導出過程の中で，情報学，複雑系との接点が見えてくる。この音声の構造的表象を用いた音 声処理系の例についても幾つか紹介し，更には，言語の起源について一つの仮説を呈する。

Abstract:
"Easiness of speech perception against variability of speech acoustics."
This is the classical problem of speech science and engineering, which remains to be solved. After the phonemes were found to be lacking in their acoustically invariant features, speech science has pursued the solution in studies of the articulation (motor theory). Speech engineering, on the other hand, has pursued it in studies of statistical modeling of speech acoustics (hidden Markov models) using large speech corpora. The problem, however, is yet to be solved. In this talk, a completely different approach is taken, where a very classical discussion in linguistics, structural phonology, is implemented physically based on a mathematical theorem of the invariant structure. This approach derives structural representation of speech, called speech Gestalt. The representation is described from viewpoints of informatics and complexity. Some examples of speech applications based on the new representation are shown and a hypothesis of the origin of language is introduced.